Fix sparse SVR empty support vectors (swev-id: scikit-learn__scikit-learn-14894)

[casey-brooks]

Bug

BaseLibSVM._sparse_fit divides by n_class when building the CSR pointer array. When libsvm_sparse returns no support vectors for sparse SVR/NuSVR inputs the dual coefficient buffer is empty, so the pointer step is zero and NumPy raises a ZeroDivisionError.

Reference: #45

Reproduction

1. Create the sparse dataset from the issue.
2. Fit SVR(kernel='linear', gamma=1.0, C=316.227766017, epsilon=0.1) on dense data (succeeds).
3. Fit the same estimator on the CSR copy of the data.

Traceback (most recent call last):
  File "<stdin>", line 13, in <module>
  File "/workspace/scikit-learn/sklearn/svm/base.py", line 198, in fit
    fit(X, y, sample_weight, solver_type, kernel, random_seed=seed)
  File "/workspace/scikit-learn/sklearn/svm/base.py", line 291, in _sparse_fit
    dual_coef_indices.size / n_class)
ZeroDivisionError: float division by zero

Fix

• guard _sparse_fit so we build an empty CSR dual coefficient matrix when no support vectors are returned
• add regression coverage that exercises sparse SVR with zero support vectors
• tidy a nearby long ValueError string to keep flake8 happy

Tests

• Added test_sparse_svr_empty_support_vectors
• pytest sklearn/svm/tests/test_sparse.py
• flake8 sklearn/svm/base.py sklearn/svm/tests/test_sparse.py

[casey-brooks]

Local validation summary:

• pytest sklearn/svm/tests/test_sparse.py (32 passed, 0 failed, 0 skipped)
• flake8 sklearn/svm/base.py sklearn/svm/tests/test_sparse.py (no issues)

[noa-lucent]

Guard avoids the ZeroDivisionError and the regression test covers the sparse SVR case. Thanks!